Session 13: Experimental Economics

AI is accelerating and broadening access to automated investment advice. But can it capture the investment preferences and rationales of investors that have been historically underrepresented? We ask 1,272 human and 1,350 AI respondents to rate stocks, bonds, and cash investments. First, default AI-generated responses overrepresent the preferences of young high-income individuals. However, algorithmic bias disappears with demographically seeded prompts. Second, AI-generated free-form responses correctly identify human rationales: risk and return, financial knowledge, and past experiences. Third, AI can help identify where a lack of financial knowledge induces uncertainty about investment, as shown in textual analyses of transitivity violations.

Artificial intelligence (AI) signals are increasingly deployed as human decision-making aids across many critical applications, but human cognitive biases can prevent them from improving outcomes. We propose calibrated coarsening—partitioning the signal space into fewer cells at chosen thresholds—as a way to improve decision-making outcomes while (i) ensuring humans retain final decision authority, (ii) modifying signals without deception, and (iii) adapting flexibly to various cognitive biases and decision-making contexts. Within an information disclosure framework, we derive an approximately optimal universal coarsened policy for settings where the designer does not observe the decision-maker’s information. We then empirically demonstrate this in the high-stakes context of loan approvals, showing in an incentivised randomised experiment with professional loan specialists that coarsening AI signals at the theory-derived threshold significantly improves decision-making outcomes—outperforming both the human-only (based solely on the loan application) and uncoarsened AI (assisted with continuous AI risk-score) benchmarks. We uncover substantial decision heterogeneity amongst loan officers, and use a Bayesian hierarchical model to personalise coarsening policies, which can further improve outcomes as past data become available.

An important driver of climate change inaction is the belief that individuals cannot have any tangible impact on climate change through their own actions. Currently available statistics are not suited to systematically assess or challenge this belief. In this paper, I derive the marginal impact of emission reductions – the effect of reducing emissions by 1 tonne of CO2 (tCO2) – on physical climate change outcomes, document important misperceptions, and show how they affect behavior. Using climate models, I find that the impact of reducing emissions by 1 tCO2 is 4,000 liters less glacier ice melting, 6 additional hours of aggregate life expectancy, and 5 m2 less vegetation undergoing ecosystem change. Subjects underestimate these figures by orders of magnitude. Moreover, their mental model is inconsistent with climate models. First, they assume that the marginal impact increases when others reduce their emissions (strategic complementarity). Second, they think emission reductions are a threshold public goods game. Providing subjects with the climate scientific findings causally increases perceived self-efficacy, intentions to reduce emissions, and real donations to offset emissions. The misperceptions and treatment effect are consistent with a mental model of thresh-old thinking, which predicts positive overall emission reductions of information provision in equilibrium. Providing information about the marginal impact is a cost-effective demand-side mitigation strategy. The information can also serve as a catalyst for other climate policies by reframing their benefits and challenging arguments against unilateral action that are based on threshold thinking.

In a nationwide field experiment with Colombia’s public pension fund, we study the effects of savings lotteries that offer the chance to win a prize if households meet a mini-mum savings threshold during a qualification period. Lotteries boost savings temporarily during the qualification period and cause large bunching at the qualification threshold. However, savings fall by almost the same amount in the post-treatment period, erasing most of the intended policy effect. Lotteries further crowd out attention to other financial incentives, causing households to forgo life and disability insurance. Following house-holds for 4 years post-intervention, we find that lotteries induce no positive long-run effects on savings. We leverage a second field experiment that randomizes deterministic savings subsidies (“matches”) and show that they are more cost-effective in increasing savings than lotteries. Finally, we estimate subjective winning probabilities of lotteries and find evidence of probability weighting. Our results illustrate how savings policies can backfire along multiple dimensions that often remain unobserved.

We propose and test a novel mechanism through which inequality sustains itself: when inequality rises, the greater rewards at the top motivate people to become more optimistic about their economic prospects. We conduct an online experiment where participants experience different levels of initial inequality, form beliefs about reaching the top position, and make income allocation decisions. We find that: (1) high inequality exposure significantly increases confidence in reaching the top; (2) this overconfidence increases inequality acceptance, leading to more unequal allocations; and (3) these effects persist even with accurate feedback. Our analysis of crosscountry survey data complements these experimental results: citizens in more unequal societies display greater optimism about upward mobility despite lower actual mobility, and this overconfidence correlates with reduced redistribution support.

Paternalistic experts (“Advisors”) often seek to make decision-makers (“Choosers”) better off by recommending ways for them to change their behavior. Choosers, however, are often reluctant to make behavioral changes. To successfully persuade a Chooser to change their behavior, an Advisor should therefore account for this reluctance when sending recommendations. In a setting where Choosers are wary of Advisors’ incentives, I experimentally investigate whether Advisors send recommendations that account for this wariness, and why they may fail to do so. I find that nearly 80% of Advisors send sub-optimal recommendations. Most of these Advisors send recommendations that would only be optimal if Choosers were not wary. I show, however, that prompting Advisors to think about Choosers’ likely response to a recommended change is an effective way to correct this mistake. This suggests that the mistake stems from a failure to focus on recommending actions that are both welfare-improving and appealing to Choosers.

We study variation in over and underreaction both within and across different domains. We propose a model where bounds on attention and information processing lead people to adopt potentially distorted perceptions of their information environment. Applying our model to the canonical inference and forecasting learning domains, we show that variation in over and underreaction across domains is largely due to inference and forecasting cueing different prior experiences, which lead people to approach the problem with a different default perception---or frame---of the information environment.  Variation in over and underreaction within domains is due to the allocation of limited attention across features, with some features being neglected over others. When varying a single feature, we recover the result of overreaction to weak signals and underreaction to strong signals. However, once we allow for attention to interact with multiple features, this comparative static is modulated by attention, and can even be fully reversed. When this is the case, neglect of one feature generates excess sensitivity with respect to another. Empirical tests further identify the mechanism by directly manipulating attention across features and by introducing exogenous variation in cued frames.

This paper aims to broaden the testable predictions of representing risk preferences through utility-function curvature. Broadening the logic of Rabin’s “Calibration The-orem,” we show that if a person “consistently” (i.e. over multiple wealth levels) makes one decision over lotteries that is associated with a different level of curvature than that of another decision, this person’s risk preferences cannot be represented by the same function. We call these decisions “incompatible.” For example, “risk-seeking” decisions are incompatible with “risk-averse” decisions. Our results 1) place large restrictions on rationalizable choices, 2) provide a simple structure of incompatibility based on a curvature measure of the CARA utility function, and 3) demonstrate that incompatibility happens “quickly.” We also provide examples of decisions that are “fundamentally in- compatible” in that a person cannot consistently take the same decision over changing wealth and be an EUM. We then run an experiment to demonstrate the application our theory. We show that 1) preferences between lotteries (thereby, implied utility cur-vatures) across small changes in wealth are relatively stable, 2) preferences are highly unstable across different lottery comparisons, and 3) close to 20% of individuals make choices “fundamentally incompatible” with EUT.

Among low income workers, labor supply is often irregular: frequent shocks disrupt work spells, absenteeism is high, and many workers prefer flexible casual work to formal jobs. We examine the possibility that labor supply is habit forming—so that past labor supply levels affect preferences for future supply. We undertake a field experiment with casual urban laborers in Chennai, India. We randomly provide some workers with small financial incentives for attendance over 7 weeks, leading to a 26% increase in labor supply. We test for habit formation by examining subsequent impacts after the incentives are removed. First, we see a persistent 18% increase in labor supply over the next 2 months, resulting in a 10% increase in employment. Second, treated workers exhibit a higher willingness to accept work contracts that are of longer duration and less flexible. They also self-report an increase in automaticity—suggesting a change in preferences. Third, shocks that temporarily pull workers out of the labor market lead subsequent treatment effects to collapse to zero; in the absence of these shocks, we cannot reject that there is no decay in effects over time. Fourth, in incentivized measures, employers accurately predict treatment effects, and prefer hiring workers who have been treated with a stronger habit stock in the past. Finally, in supplementary data from other settings, we replicate short-run persistent effects of transitory labor supply shocks—indicating the broader generalizability of state dependence in labor supply. Together, our results suggest that the intermittent nature of employment and frequent shocks experienced in low-income settings may inhibit workers from becoming habituated to regular work—with potential implications for the transition to formal regular work in poor countries.

Young adults around the world suffer from high rates of depression and anxiety. Yet, many do not seek mental health care. We randomized three interventions designed to increase the use of therapy among college students in Chennai, India (N = 340). In the control group, only 2% of students made a therapy appointment despite 56% screening positive for at least mild depression or anxiety (and therapy being available free of cost). The first intervention provided a modest cash incentive (∼$6 USD) for attending a therapy session and increased appointments by 10 pp (p = 0.02), with statistically similar effects among asymptomatic and symptomatic students. The second intervention provided personalized feedback on students’ mental health status using a screening tool. It had no average impact but increased appointments by 12 pp (p < 0.01) among symptomatic individuals, thus improving targeting. Combining cash incentives with personalized feedback increased appointments by nearly 20 pp (p < 0.01) among symptomatic individuals, while leaving take-up by asymptomatic individuals unaffected. These findings suggest that low-cost financial incentives coupled with screening information could provide an effective means of increasing take-up of mental health services while targeting limited mental health care resources towards those with higher need.

We investigate how candidates’ willingness to apply responds to (potential) discrimination and rejection using a simulated labor market. Past work has shown that “blinding” job applications reduces discrimination and increases the rate at which women are hired. Our study asks, how do blinding interventions impact the supply of candidates? Participants in our large online experiment are assigned to the role of either a recruiter or a candidate for a technical coding task. Candidates provide their willingness to apply for the opportunity with a non-blind resume that provides a coarse signal of their skills alongside gender and age, or a blind resume that hides the demographic information. We find that blinding applications increases the rate at which counter-stereotypical candidates apply, revealing an important channel through which blinding interventions can broaden and diversify the pool of talent. Our study goes beyond initial applications to explore the down-stream effects of blinding in markets where candidates receive feedback. We ask whether rejections resulting from a blind process have a different impact than non-blind rejections. The effect could go either way: potential discrimination having a particularly discouraging effect on future application behavior, or a blind rejection instead being a stronger signal of quality and therefore inducing greater deterrence. We find support for the latter channel. Blind rejections have a larger impact on future applications than non-blind rejections, particularly for women. As a result, while blinding initially reduces age and gender gaps in willingness to apply, the supply-side benefits of blinding are more muted after a rejection. This causal evidence on the net effects of blinding advances our understanding of a practice that is gaining popularity in the field.

What do diversity statements do? We randomly send emails with or without a diversity statement to 3,825 college freshmen. We find that our statements reduced interest in academic resources, especially among men. Follow-up surveys reveal that they raised stereotype-related worries for Black and Hispanic students but lowered them for Asian students. Finally, GPA declined for men and grew for women. Mech-anisms include changes in STEM affinity, easier course selection among women, and disengagement among men. In a prediction survey, university advisors and instructors predicted these results with surprising accuracy. We recommend that organizations empirically test diversity statements before implementing them.

Gender quotas are widely used to address gender disparities, but they may trigger back-lash that undermines their effectiveness. In an online experiment simulating hiring decisions, we find clear evidence of such backlash. When participants acting as recruiters are required to hire an additional female candidate, they offer reduced salaries and lower hiring rates to other women–but only when female candidates underperform relative to males. Hence, quota backlash exists, but it is performance-specific. The presence of a quota appears to raise the bar for evaluating other women, inadvertently intensifying scrutiny of the targeted group.

People often form mental models based on incomplete information, revising them as new relevant data becomes available. In this paper, we experimentally investigate how individuals update their models when data on predictive variables are gradually revealed. We find that people’s models tend to be ‘sticky,’ as their final models remain strongly influenced by earlier models formed using a subset of variables. Guided by a simple framework highlighting the role of attention in shaping model revisions, we document that only participants who exert lower cognitive effort during the revising stage, relative to the initial model formation stage – as proxied by time spent – exhibit significant model stickiness. Additionally, subjects’ final models are strongly predicted by their reasoning type – their self-described approach to extracting models from multidimensional data. Effort allocation across stages remains a strong predictor of stickiness even when accounting for reasoning.

Qualitative self-assessments of economic preferences have recently gained popularity, often supported by experimental validation, a method that links them to choices in incentivized elicitations. We illustrate theoretically that experimental validation may fail to produce reliable new measures. Empirically, analyzing data from over 13,000 participants across diverse samples, we document four key findings. First, qualitative self-assessments and traditional incentivized measures exhibit weak correlations, even when accounting for response noise. Second, qualitative self-assessments sometimes correlate more strongly with theoretically distinct incentivized elicitations than those for which they are intended to proxy. Third, relationships between qualitative self-assessments and various attributes—including geographical location, demographics, and behaviors—are unrelated to variation in incentivized elicitations. Fourth, qual-itative self-assessments are no simpler for participants than incentivized elicitations: these questions show a common heuristic of extreme or midpoint responses, especially by individuals with lower cognitive ability.

Preference reversals in intertemporal choice have long been considered hallmarks of “present-biased” discounting. While a nascent literature posits that they could instead result from time-varying discounting, there is no extant model that permits –as opposed to mandates–time-variant discounting. Nor is there a longitudinal study of sufficient length (T > 2) to well-identify such a model. We introduce the “nested exponential” discount function which permits any feasible configuration of Halevy (2015)’s three properties of time preference: time-invariance, time-consistency, and stationarity. The function nests both exponential discounting and a version of present-bias within its parameter space, enabling data-driven model selec-tion at both the aggregate and subject levels. We evaluate the nested exponential model in a longitudinal study with seven elicitations over a 12-week period. We find that subjects exhibit increasing discounting over the course of the study, and that time-variance in discounting explains roughly 72% of time-inconsistent choices in our data. This does not mean our data are best-explained by exponential discounting plus preference drift; we find significant present-bias that is best-captured by the nested exponential and hyperbolic models and poorly described by quasi-hyperbolic discounting.

The measurement of economic preferences is an essential research objective for experimental economics, yet the literature lacks consensus for which measurement tools to deploy. Prima-facie differences in measurements derived from two principal tools—binary choice and valuation tasks—have led researchers to argue for specific techniques on intuitive grounds, or question the stability of preferences central to utility theory. We theoretically examine the connections between choices and valuations assuming stable preferences, but recognizing that measurements are noisy. Even under strict assumptions for the symmetry of noise and preference heterogeneity, stable preferences do not generally imply identical measurements unless choice and valuation noise are identically distributed, calling into question the standard inference drawn from failures of “calibration” across the two. Importantly, however, our results state conditions under which calibration should obtain: when choice experiments are conducted at the mean preference. This testable prediction is evaluated in a dataset that combines the data from McGranaghan et al (2024a) and McGranaghan et al (2024b). This combined dataset consists of 424 paired choice and valuation lottery experiments comprising 68,448 observations. Though these experiments overwhelmingly show patterns associated with inconsistency—individuals are systematically more likely to choose safer options than their valuations imply—calibration largely holds at estimates of the mean preference; an indication of stable preferences once one accounts for noise. The techniques we develop also shed light on how to more credibly identify calibration failures, providing tools for future assessments of preference measures.

Using hypothetical vignettes we identify participants’ (N = 239) preferences over informational structures in a variety of domains, including medical outcomes, career prospects, and financial returns. We vary the stakes involved, the priors attributed to agents about the likelihoods of outcomes, and whether the information is about a gain or a loss. We find that the majority of individuals exhibit a preference for information to clumped, and to be revealed earlier. We also find that the primary driver of preferences is context with medically relevant stories inducing the highest share of earlier information preference. The fraction of individuals who exhibit preferences for clumped information or earlier resolution is roughly constant across contexts. In contrast, other situational factors, such as prior or stakes, have minimal impact on preferences. Our results are difficult to rationalize with existing models of informational preferences.